Although the rehabilitation of aphasia has been extensively studied, the prediction of language outcome still has not received sufficient attention. The aim of this study was to predict the language outcome using mismatch negativity (MMN) in patients with large left-hemispheric infarction. Methods: MMN was elicited by an oddball paradigm in which a standard tone (1000 Hz) and deviant tone (1500 Hz) were presented at 90% and 10% of the number of tones, respectively. The mean amplitudes and laterality indexes (LIs) of MMN were measured over the prefrontal, frontal, central, parietal, temporal, and perisylvian electrodes and both hemispheres during the first 7 days (session 1) and 10 to 20 days (session 2) post-onset. Mixed three-way analysis of variance (ANOVA) was used to investigate differences in these factors between two aphasia groups (the good recovery group and poor recovery group). The predictive value of the most significant LI was also compared with the score of National Institutes of Health Stroke Scale score and low-density volume on computed tomography. Results: A total of 18 patients were enrolled in this study. Mixed three-way ANOVA showed no interaction effect of session × region of interest (ROI) × group (F [3.59, 57.38] = 1.301, P = 0.282) and no interaction effect of ROI × group (F [1.81, 29.01] = 0.71, P = 0.487) and session × group (F [1.00, 16.00] = 0.084, P = 0.776) for MMN amplitude. No interaction effect of session × ROI × group (F [1.79, 28.58] = 0.62, P = 0.530), but an interaction effect of session × group (F [1.00, 16.00] = 5.21, P = 0.036) was found for LIs. In the poor recovery group, the LIs of MMN over all the ROIs, except the parietal area, became more negative at session 2 than those at session 1 (P < 0.05), but this effect was not observed in the good recovery group. Additionally, significant differences were observed in the LIs at session 2 between the two groups (P < 0.05). The LI over the perisylvian area at session 2 had the highest predictive value with an area under the curve of 0.963 (95% confidence interval: 0.884-1.000). An LI score >-0.36 over the perisylvian area suggested good recovery, but a score <-0.36 suggested poor recovery. The LI cut-off value of -0.36 had the highest sensitivity (90.0%) and specificity (87.5%) for predicting a good language outcome at 3 months post-stroke. Conclusion: LIs of MMN amplitudes a approximately 2 weeks post left-hemisphere stroke serve as more sensitive predictors of language outcome, among which the LI over the perisylvian area exhibits the best predictive value. © 2019 The Chinese Medical Association, produced by Wolters Kluwer, Inc.